Agent Lightning v1.0: Toward Harnessed Agentic RL

Modern AI agents operate within an Agent Harness that manages tools, context, and control flow, making it a core component of agent systems. The existing Agent Lightning introduced a distributed architecture that connects arbitrary agents to RL training via an LLM endpoint proxy, evolving this approach into a new paradigm called Harnessed Agentic RL.

Harnessed Agentic RL is fundamentally different from traditional agent RL. The environment interaction loop is owned by the deployment-time harness rather than the training engine, and the trainer only observes sequences of LLM request-response pairs. This gives rise to new technical challenges affecting training stability and effectiveness, such as retokenization, sample merging, advantage calculation, loss normalization, and backend scheduling.

Agent Lightning v1.0 is a lightweight framework implemented in approximately 3,500 lines of code, serving as a practical testbed for studying these challenges. It supports arbitrary agent harnesses and has conducted evaluations on instruction-following, retrieval, and coding agents. In particular, it provides a fully reproducible pipeline for coding agent RL, achieving a 14.6 percentage point (absolute) improvement for the Qwen3.5-9B model on the SWE-bench Verified benchmark, from 41.8% to 56.4%. This result was achieved with only 6K training examples and modest computing resources.